Oracle Corporation

01 — Executive summaryRevenue of $19.3 billion, a backlog of $664 billion, and free cash flow of negative $5.4 billion. All three are correct.

Oracle's first fiscal quarter of 2027, reported on 10 September 2026, was the best operating quarter in the company's sixty-year history by almost any measure that describes demand. Revenue rose 29.6% to $19.345 billion. Cloud infrastructure revenue more than doubled, rising 121% to $7.4 billion. The remaining performance obligation — the dollar value of signed contracts not yet recognised as revenue — reached $664 billion, an increase of $209 billion in twelve months and $26 billion in three. Oracle delivered 850 megawatts of new data centre capacity and more than 300,000 graphics processors in the quarter. Non-GAAP operating margin held at 42.1%. GAAP operating income rose 57%.

The same quarter also produced $5.4 billion of negative free cash flow, $1.4 billion of interest expense, and a further $28.5 billion of capital expenditure against $23.1 billion of operating cash flow — of which $11.4 billion was customer money received in advance and booked as an operating inflow. The company carries $125.3 billion of financial debt and $30.6 billion of operating lease liabilities. It holds a BBB− rating from S&P Global Ratings, one notch above speculative grade. Its five-year credit default swap spread briefly exceeded the level it reached at the peak of the 2008 financial crisis. And the shares, at $139.54, sit 56.7% below their 52-week high.

Both paragraphs describe the same company in the same quarter. The investment question is not which one is true. It is which one is load-bearing, and how much of the second is the mechanical consequence of the first.

On the first, the disclosed picture is uncomfortable but improving at the margin. S&P cited OpenAI by name as a key credit risk when it downgraded Oracle in July 2026, and reported analysis places OpenAI at roughly half of the backlog. Management said on the September call that the non-OpenAI backlog more than doubled over the past year, which is a genuine diversification of the book even if the concentration remains material. OpenAI lost $20.9 billion in 2025 and funds itself through successive equity rounds; its ability to pay a $300 billion compute bill is a function of whether those rounds continue.

On the second, the arithmetic is knowable and unfavourable in the near term. Oracle depreciates servers and networking equipment over six years. Gross capital expenditure is running at $92.5 billion a year against quarterly depreciation of $3.2 billion, or roughly $12.6 billion annualised. The depreciation charge implied by a steady-state programme of this size is therefore an order of magnitude larger than the charge recognised today, and it will arrive over the next three years whether or not the revenue does. Gross margin has already fallen from 79.1% in fiscal 2022 to 64.0% on a trailing basis. Operating margin has risen over the same period only because operating expenses have been held flat while revenue grew.

None of this makes Oracle a bad business. It makes it a different business from the one the multiple historically described. Oracle was a 70%-gross-margin software annuity trading on cash generation. It is becoming a capital-intensive infrastructure operator whose economics depend on utilisation, pricing power and the cost of capital. Those are good businesses when run well. They are not valued the way licence software was valued, and the market has spent the last twelve months working that out.

of contracted backlog against $71.8 billion of trailing revenue — 9.2 times coverage, of which roughly 12% is scheduled to convert inside twelve months.

Our scenario framework produces a range of $59 to $290 per share, with a probability-weighted value of $159. We apply a 30% weight to a bear case in which AI infrastructure demand normalises before the capacity is monetised and the depreciation burden compresses margins; a 50% weight to a base case in which the backlog converts broadly on schedule and operating margin settles in the high thirties; and a 20% weight to a bull case in which Oracle's cost position in power and land lets it earn infrastructure margins above the industry. Our 12-month target is $160, approximately 15% above the current price. We rate the shares Neutral.

02 — Company & business modelThree businesses, one balance sheet, and a fourth that did not exist three years ago

Oracle Corporation was founded in 1977 and built its franchise on the relational database. For four decades the model was straightforward and extremely profitable: sell perpetual licences to large enterprises at very high gross margin, then charge an annual support fee of roughly 22% of the licence value, which renews at close to 100% because the cost of migrating a running database is prohibitive. That support annuity is the reason Oracle's revenue was, for most of its history, unusually insensitive to the economic cycle.

The company has been migrating that annuity into the cloud since 2016. What is new is the scale and the character of the migration. Through fiscal 2024, Oracle Cloud Infrastructure was a credible but modest third or fourth player: a platform for Oracle's own database workloads, growing quickly from a small base, and not a business that required a different balance sheet. Between the quarter ended 31 May 2025 and the quarter ended 31 August 2026 that changed completely. Cloud infrastructure revenue went from a $10.2 billion annual run-rate to a $29.6 billion one. Property, plant and equipment, net, went from $43.5 billion to $127.8 billion. Debt went from $92.6 billion to a peak of $134.6 billion and back to $125.3 billion. Total assets went from $168.4 billion to $303.3 billion.

The revenue mix in the chart above is the most important table in this report, and it is the one most often misread. The largest single pool is still the legacy software line — licence and support — at $24.5 billion, or 36.4% of fiscal 2026 revenue. It is also the only line that is shrinking, down 1% for the year and down 3% in the most recent quarter. Cloud applications, at $16.0 billion, is the stable, high-margin, slow-growth business that most investors imagine when they think of Oracle. Cloud infrastructure, at $18.0 billion in fiscal 2026, is the new business, and it is the one that consumes essentially all of the capital.

That produces a structural asymmetry worth stating plainly. Roughly 45% of revenue — software plus cloud applications — behaves like a mature, high-margin, capital-light annuity. Roughly 27% — cloud infrastructure — behaves like a utility: capital-hungry, price-competitive, and margin-dependent on utilisation. The remaining 13% is services and hardware, which are lower-margin and largely derivative of the other two. The company's consolidated margin is therefore a weighted average of two businesses with very different economics, and the weight is shifting fast toward the lower-margin one.

Oracle is also, functionally, a healthcare company and a construction company now. Oracle Health — the former Cerner business, acquired for $28.3 billion in 2022 — generated a mid-single digit share of revenue and has been a margin drag since acquisition. And the cloud infrastructure business has made Oracle one of the largest private purchasers of power, land and construction services in the United States, which brings a set of risks that do not appear in a software company's risk factors. We cover both in sections 18 and 22.

Governance sits with a co-CEO structure. Clay Magouyrk, who built OCI, and Mike Sicilia, who ran the applications and industry businesses, were appointed co-CEOs in September 2025 when Safra Catz moved to the executive vice-chair role. Hilary Maxson, formerly of Schneider Electric, is chief financial officer. Larry Ellison remains executive chairman and chief technology officer and, at approximately 40% of the equity, retains effective control of the company's direction. That concentration of ownership is a genuine asset in a capital cycle that requires a five-year view. It is also a governance fact that limits the influence of outside shareholders over a strategy the chairman has described in the most expansive terms.

03 — The financial recordSix years of accelerating revenue, and a margin structure that has quietly inverted

Oracle's fiscal 2026, ended 31 May 2026, produced record revenue of $67.357 billion, up 17.4%. GAAP operating income was $20.606 billion and GAAP net income was $16.984 billion, or $5.83 per diluted share. Non-GAAP earnings per share were $7.63 including one-time investment gains and $6.83 excluding them. Operating cash flow was a record $31.977 billion, up 54%. Capital expenditure was $55.663 billion. Free cash flow was negative $23.686 billion.

The revenue trajectory across six years is close to a straight line of acceleration: $42.4 billion in fiscal 2022, $50.0 billion, $53.0 billion, $57.4 billion, $67.4 billion, and a trailing twelve-month figure of $71.8 billion. Growth went 4.8%, 17.7%, 6.0%, 8.4%, 17.4%, and is now running at 29.6% in the latest quarter on company guidance of 30% to 34% for the December quarter. Very few companies of Oracle's size have re-accelerated from single digits to thirty percent, and the reason is unambiguous: a single product line.

The profit record is more interesting than the revenue record, because the two moved in different directions from the fourth quarter of fiscal 2026 onward. GAAP operating income grew 57.3% in the September quarter against revenue growth of 29.6%, which is operating leverage of an unusual kind. The mechanism was not pricing. It was that sales and marketing fell 12.2% to $1.811 billion, research and development fell 3.6% to $2.401 billion, amortisation of acquired intangibles fell 52% to $202 million, and restructuring fell 77% to $94 million. Total operating expenses rose 18.5% against revenue growth of 29.6%.

A company that is cutting sales and marketing by 12% in the quarter its revenue grows 30% is not investing behind a demand-led expansion. It is reallocating. Oracle has been moving engineers and go-to-market capacity toward the infrastructure business, reducing headcount in the legacy lines, and letting the applications business run on installed-base renewals. The company reported approximately 141,000 full-time employees at the end of fiscal 2026. That is fewer than it employed two years earlier, on materially higher revenue.

The margin chart is the central fact of this report. Gross margin has fallen 15.1 percentage points in four years, from 79.1% in fiscal 2022 to 64.0% on a trailing basis. Almost all of that decline is mix: cloud infrastructure revenue carries a gross margin in the low thirties, according to the company's own disclosures, against roughly 80% for licence and support and something in the sixties and seventies for cloud applications and multicloud database. As infrastructure revenue grows from 27% of the total toward something like half, the consolidated gross margin must fall further. This is arithmetic, not opinion.

Operating margin, by contrast, has risen from 27.4% in fiscal 2023 to 34.3% on a trailing basis. The reason is that operating expenses have not grown with revenue. That is sustainable for a while — there is genuine operating leverage in a platform whose development is already sunk — but it is not sustainable indefinitely, because the depreciation on $127.8 billion of net property, plant and equipment has barely begun to be charged. Section 16 works through that arithmetic in detail. The short version is that gross margin is falling for structural reasons and operating margin is rising for temporary ones.

The cash flow record is where the divergence between reported profit and economic reality is starkest. Oracle has reported positive GAAP net income in every one of the last six fiscal years, cumulatively $74.8 billion. Over the same period it has consumed $33.0 billion of free cash flow, because capital expenditure of $118.8 billion exceeded operating cash flow of $138.5 billion. In fiscal 2024 the company generated $11.8 billion of free cash flow. In fiscal 2026 it consumed $23.7 billion. On a trailing basis it is now consuming $28.7 billion a year.

04 — Segment: Cloud infrastructureThe only line that matters, growing 121% from a base that was not there two years ago

Oracle Cloud Infrastructure generated $7.388 billion of revenue in the quarter ended 31 August 2026, an increase of 121% in US dollars. On an annualised basis that is $29.6 billion, against $10.2 billion two fiscal years earlier. The segment is now the second largest revenue line in the company and will pass the legacy software line within roughly four quarters at the current trajectory.

Three things drove the quarter. First, delivery. Oracle added 850 megawatts of data centre capacity in the three months, and has now delivered more than 300,000 graphics processors to its AI cloud customers since the end of the previous quarter, nearly triple the volume delivered in that quarter. At Abilene, the flagship campus in Texas, 131,000 processors were delivered in the quarter, 1.9 times the prior quarter's volume; six of eight buildings, representing 618 megawatts and 75% of total campus capacity, have now been handed over to the customer, with acceptance compressed to 24 hours. Second, conversion. Management states that roughly half of the backlog will convert to revenue within thirty-six months, and the conversion has begun: the sequential increase in infrastructure revenue was the largest in the company's history. Third, pricing. Of all the processors that came up for renewal in the quarter, the capacity was renewed or resold at a 20% premium to the prior contract, and the majority of that hardware was four years old or more.

That renewal statistic is the single most favourable piece of evidence in the entire bull case, and it deserves to be examined rather than repeated. If four-year-old GPUs are renewing at a 20% premium, then either the market for AI compute is tight enough that old silicon retains pricing power, or the original contracts were written below market, or the comparison base is unrepresentative. Oracle does not disclose the absolute value of the renewals, so a 20% premium on a small base tells us little. What it does establish is that Oracle is not, at present, writing off capacity it built speculatively. Utilisation was 97.9%, which is close to full.

The segment's economics are less flattering than its growth. Oracle does not disclose a gross margin for cloud infrastructure specifically, but the cloud and software cost of revenue line rose 77% to $6.4 billion in the September quarter against cloud revenue growth of 62%. Costs are growing faster than the revenue they support. Management has confirmed that the infrastructure business carries a gross margin in the low thirties on AI workloads, and disclosed that multicloud database revenue carries 60% to 80% margins. The blended cloud gross margin is therefore falling as infrastructure displaces applications.

There is also a question of what the revenue actually represents. Oracle's AI contracts are structured in several different ways. In some, Oracle buys the GPUs and leases compute to the customer. In others, the customer prepays for the hardware, or supplies the hardware itself and pays Oracle for the data centre, power, networking and operations. Management disclosed in June 2026 that the prepaid and customer-supplied hardware portions of its large AI contracts total $75 billion. A "bring your own hardware" arrangement is a much thinner revenue line than a fully hosted compute lease, and Oracle does not break out the mix.

05 — Segment: Cloud applicationsGrowing 10%, carrying the margin, and doing the work nobody notices

Cloud applications — the Fusion and NetSuite suites covering enterprise resource planning, human capital management, supply chain and customer experience — generated $4.2 billion in the September quarter, up 10%. For the full fiscal year the line produced $15.975 billion, up 11%. That is a respectable growth rate for a mature application suite and a disappointing one for a business Oracle describes as strategically central.

The applications business matters to this report for three reasons that have nothing to do with its growth rate. First, it is where the margin is. Cloud applications carry a gross margin in the sixties to low seventies, several times the infrastructure business. On $16.0 billion of revenue it plausibly contributes more gross profit than the $18.0 billion infrastructure line did in fiscal 2026. Second, it is the anchor that makes the infrastructure business defensible. Oracle's structural argument against AWS, Azure and Google Cloud has never been that OCI is a better general-purpose cloud. It is that Oracle offers a complete vertical stack — database, applications and infrastructure — and that a customer running Fusion on OCI avoids the data egress and latency costs of running it anywhere else. Third, it is the source of the recurring base that funds the capital programme. The applications and support annuity produces most of the operating cash flow that the infrastructure business consumes.

The risk is that the applications business is being starved. Sales and marketing fell 12.2% year over year in the September quarter. Research and development fell 3.6%. Some of that is efficiency, and Oracle has been explicit that AI-assisted development is reducing its own engineering cost. But a 10% growth rate in a market growing faster than 10%, with declining investment behind it, is how a franchise begins to lose share. SAP, Salesforce and Workday all compete in the same suites, and all three are investing at or above their revenue growth rates.

There is also the question of the Oracle Health applications business, which we treat separately in section 18. It is a large installed base, a low margin, and a promised AI rewrite that management expects to push growth into double digits in fiscal 2027.

06 — Segment: Software, hardware & servicesThe annuity that pays the bills, and the slow leak in it

The software line — cloud licence and on-premises licence, plus licence support — was $24.541 billion in fiscal 2026, down 1%, and $5.550 billion in the September quarter, down 3%. It is still the largest single revenue pool in the company at 36.4% of the total. It is also the line with the highest gross margin, in the region of 80%, and the one that generates the cash that funds everything else.

The decline is deliberate and structural. Oracle's on-premises customers are migrating to Oracle Cloud, and when they do, licence and support revenue becomes cloud revenue at a lower margin. Management states this explicitly in every earnings release: "reflecting our customers' continuing migration from on-premises software to the Cloud." The migration is a good thing for the durability of the franchise and a bad thing for the reported margin, and it will continue until the installed base is exhausted.

The important question is the pace. Support revenue is extremely sticky: cancelling a database support contract means forgoing patches, security updates and the right to upgrade, which most regulated enterprises will not do while the database is still running production workloads. So the support line should decay slowly, at a rate determined by how quickly customers retire their last on-premises instances. A 1% annual decline is consistent with a very long tail. A 3% quarterly decline, if sustained, is not.

We treat the September quarter's 3% decline as noise rather than signal — one quarter is not a trend, and the prior-year comparative included a large one-off licence transaction. But the line is worth watching closely, because the entire capital plan assumes that the legacy annuity keeps paying while the infrastructure business scales. If support revenue begins to decline at high single digits, the funding arithmetic for the build changes materially.

Hardware, at $3.084 billion for the year and $774 million in the quarter, is the engineering systems and Exadata appliance business. It grew 15% in the quarter, which is faster than it has grown in years and reflects pull-through demand from AI customers who want Oracle's high-performance networking and storage appliances alongside the compute. At 4.6% of revenue it is immaterial to the consolidated picture but useful as a leading indicator: hardware is sold alongside infrastructure deployments, so its growth corroborates the delivery statistics.

Services, at $5.743 billion for the year and $1.414 billion in the quarter, is consulting and advanced customer support. It grew 5% in the quarter on revenue up 30%. That divergence is worth noting: Oracle is deploying enormous quantities of infrastructure without a proportionate increase in the consulting effort required to install it. Part of the explanation is that AI cloud deployments are closer to turnkey than traditional enterprise implementations. Part of it is that the integration work is being done by the customers, or by their own systems integrators, rather than by Oracle. Either way, services is not scaling with the business it is supposed to support.

Taken together, the non-infrastructure lines generated $49.3 billion of fiscal 2026 revenue, grew at roughly 5% blended, and contributed the overwhelming majority of consolidated gross profit. The infrastructure line generated $18.0 billion, grew 77%, and consumed essentially all of the capital. Understanding Oracle in 2026 means holding both of those facts at once.

07 — The multicloud databaseThe fastest-growing business Oracle has ever had, and the one nobody talks about

Oracle's multicloud database business grew 353% year over year in the September quarter, and 404% in the June quarter before it. Management called it, in the June release, "our fastest growing business ever." It is also the least understood part of the company, and in our view the most strategically important.

The product is straightforward to describe. Oracle Database is the market's dominant enterprise relational database, running a large share of the world's transaction processing, and it has historically only run well on Oracle's own infrastructure. The multicloud offering puts Oracle Database inside the data centres of Microsoft Azure, Amazon Web Services and Google Cloud, so that an enterprise can keep its application stack on whichever hyperscaler it prefers while continuing to run Oracle Database underneath it. The customer gets a single database contract and a single support relationship across four clouds.

The strategic logic is that it removes the principal objection to Oracle's database franchise. For a decade, the reason enterprises migrated off Oracle Database was not price or capability; it was that running Oracle meant committing to Oracle's cloud, and many large customers had already committed to somebody else's. Multicloud converts a competitive disadvantage into a distribution advantage: Oracle stops fighting the hyperscalers for the primary cloud relationship and instead sells into their estates. Multicloud customer count is up 180%.

The economics are the best in the company. Management disclosed in March 2026 that the multicloud database business carries gross margins between 60% and 80%, which is several times the margin on AI infrastructure and comparable to licence software. It requires no material incremental capital, because the compute belongs to the partner. And it is a subscription business with the same switching costs as the legacy annuity, because migrating a production database is a multi-year project.

The caveat is competitive. AWS, Azure and Google Cloud host Oracle Database because their customers demand it, not because they want to. Each of the three has spent a decade and enormous sums building a competing database — Aurora, Cosmos DB, Spanner, AlloyDB — and each has a commercial interest in migrating Oracle workloads onto it. The multicloud arrangement gives the hyperscalers exactly the proximity to Oracle's installed base they need to compete for it. Oracle is being paid to help its competitors displace it, and the question is whether the cash arrives faster than the displacement.

08 — The backlog$664 billion of contracted revenue, and the step change that produced it

Remaining performance obligations are the dollar value of contracted work that has not yet been recognised as revenue. At 31 August 2026 Oracle reported $664 billion, up $209 billion year over year and $26 billion sequentially. That is 9.2 times trailing twelve-month revenue, an extraordinary figure for any company and one without a close precedent in enterprise software.

The chart above is not a growth curve; it is two different regimes separated by a single quarter. Through the first four quarters shown, Oracle's backlog moved between $97 billion and $138 billion, growing steadily and unremarkably. In the quarter ended 31 August 2025 it jumped from $137.8 billion to $455.3 billion — a $317.5 billion increase in three months, larger than the company's entire revenue base at the time. The cause was the signing of a small number of very large AI infrastructure contracts, the largest of which was the reported $300 billion, five-year agreement with OpenAI announced in September 2025.

It is worth being precise about what the backlog does and does not mean. It is not a forecast of revenue. It is a contractual commitment, which is a stronger thing, subject to three qualifications that matter enormously here. First, timing: the contracts are long, and the revenue arrives on a schedule the customer controls, not Oracle. Second, credit: a contract is only worth what the counterparty can pay. Third, cancellation: some large infrastructure agreements include termination provisions, and Oracle does not disclose them in detail.

On timing, management has given one useful data point: roughly half of the backlog is expected to convert to revenue within thirty-six months. Sell-side work on the disclosed ramp suggests approximately 12% converts inside twelve months and a further 34% in the following twenty-four. That means roughly $80 billion of the $664 billion becomes revenue in the next year — about 12 cents on the dollar — and roughly $300 billion over three years. The remaining $330 billion sits beyond 2029.

This is not a defect in the contracts. It is what a five-year, staged data centre buildout looks like when it is expressed as a single balance sheet number. A five-year contract cannot recognise most of its value in year one regardless of how healthy the buyer is. The problem is that the market treats a $664 billion backlog as though it were a $664 billion order book, when it is closer to a $300 billion three-year order book attached to a $364 billion option on continued demand.

There is one further mechanical point that bears on the quality of the number. Backlog grows when contracts are signed and shrinks when revenue is recognised. Oracle signed more than $30 billion of new AI cloud contracts in the September quarter while recognising $7.4 billion of infrastructure revenue, so the net movement was strongly positive. But the company has said that the bulk of the newly signed contracts will not affect revenue until fiscal 2028 or later. A backlog that grows because new contracts are long-dated, while the revenue from old contracts arrives slowly, is a backlog whose average duration is lengthening. That is not the same as a backlog that is converting.

09 — Quality of the backlogHalf of it belongs to one counterparty that has never made a profit

When S&P Global Ratings downgraded Oracle to BBB− on 9 July 2026, it cited OpenAI by name as a key credit risk and noted that the AI developer represents roughly half of the company's backlog. The downgrade was not about Oracle's operating performance, which was improving. It was about counterparty concentration.

The arithmetic is stark. If roughly half of $664 billion is attributable to OpenAI, that is approximately $332 billion of contracted revenue from a single private company. OpenAI has never reported a profit. It lost $20.9 billion in 2025 on an operating basis, and its capacity to pay a $300 billion compute bill depends on its ability to keep raising equity. It closed a $110 billion round in February 2026 — Amazon contributing $50 billion, NVIDIA and SoftBank $30 billion each — at a valuation reported near $840 billion. Each new round is evidence that the payment capacity exists. A stalled round would be evidence that it does not.

The counter-argument, which management made explicitly on the September call, is that the book is diversifying. The non-OpenAI backlog more than doubled over the past year. New bookings in the quarter exceeded $30 billion and were spread across a wider customer set including NVIDIA, Meta, Advanced Micro Devices, xAI and TikTok's parent. If the non-OpenAI book is growing faster than the OpenAI book, concentration is falling even as the absolute exposure rises. That is a real improvement and it should be credited.

But it is worth observing what kind of diversification this is. NVIDIA, AMD, Meta and xAI are not independent counterparties in the ordinary sense. NVIDIA and AMD are Oracle's suppliers of the compute it resells. Meta and xAI are themselves building competing infrastructure. Several of the largest customers are simultaneously suppliers, competitors and, in some cases, investors in each other. When the same pool of capital finances the supplier, the customer and the landlord, the credit quality of the customer is not independent of the landlord's continued ability to build. That circularity is not fraud and it is not necessarily unsound. It does mean that a stress test which assumes the counterparties fail independently will understate the correlation.

There is a final disclosure gap. Oracle does not name its largest customers in the backlog discussion, does not disclose the concentration percentage in its filings, and does not break out the segment margin on AI infrastructure. The S&P estimate of roughly half is the best public number available and it comes from a rating agency, not from the company. An investor cannot verify it, and in our view the absence of the disclosure is itself informative: a company confident in the diversification of a $664 billion book would publish the split.

10 — Customer prepaymentsThe most important number in the quarter, and the one that is easiest to miss

Oracle reported record operating cash flow of $23.103 billion in the September quarter, up 184% year over year. The press release led with it. Buried in the condensed consolidated statement of cash flows, inside the changes in operating assets and liabilities, is the line that explains most of it: an $11.363 billion increase in deferred revenues arising from "customer prepayments with significant financing component."

That line did not exist a year earlier, when the comparable figure was zero. It represents cash received from customers today for capacity to be delivered years from now. It appears in the operating section of the cash flow statement, and under US generally accepted accounting principles that treatment is defensible — the cash is collected from a customer, not a lender. But economically it is a financing inflow. Oracle is borrowing from its customers, interest-free or at an embedded discount, and the money is being used to buy the assets that will generate the revenue.

Adjusting for it changes the picture materially. Core operating cash flow was $11.7 billion, not $23.1 billion. Against $28.5 billion of capital expenditure, the true free cash flow deficit was $16.8 billion, not the reported $5.4 billion. Both numbers are correct; they answer different questions. The reported figure answers "how much cash moved through the operating section." The adjusted figure answers "how much cash did the business generate from selling things."

The balance sheet corroborates the flow. "Other non-current liabilities" rose from $16.178 billion to $28.183 billion in the quarter, an increase of $12.0 billion, which is where long-dated customer prepayments sit. Current deferred revenues rose from $9.916 billion to $14.686 billion. Together, the two lines show roughly $16.8 billion of customer money received in three months. Oracle's disclosure that prepaid and customer-supplied hardware across its large AI contracts totals $75 billion is the cumulative figure.

The capital programme chart shows how this works in the guidance. Gross capital expenditure is guided to $90 to $95 billion in fiscal 2027. Net cash capital expenditure — the amount Oracle itself has to find — is guided to no more than $70 billion. The difference of roughly $22.5 billion is customer-funded. We derive that figure by subtraction; Oracle does not disclose it as a line, and the company has been careful to distinguish "no additional cash from Oracle" from "no additional capital expenditure."

That distinction matters more than it appears. A customer that prepays for hardware still requires Oracle to order it, install it, power it, network it and operate it. The asset may sit on the customer's balance sheet or on Oracle's, but the operational complexity, the construction risk, the power contract and the execution risk all sit with Oracle. Prepayment transfers funding risk. It does not transfer delivery risk, and delivery risk is where the Project Jupiter force majeure notice of 24 September 2026 becomes relevant — a point we return to in section 22.

11 — Capital intensity$92.5 billion a year, thirteen times the fiscal 2024 figure, on revenue that grew 27%

Oracle spent $6.866 billion on property, plant and equipment in fiscal 2024. It spent $21.215 billion in fiscal 2025, $55.663 billion in fiscal 2026, and has guided to $90 to $95 billion in fiscal 2027. In the September quarter alone it spent $28.499 billion, which is more than the whole of fiscal 2025 and 4.1 times the prior-year quarter.

Over the same four-year period, revenue grew from $52.961 billion to a guided $90 billion, an increase of 70%. Capital expenditure grew thirteenfold. The ratio of capital expenditure to revenue has gone from 13% in fiscal 2024 to 40% in fiscal 2026 and something like 103% of guided fiscal 2027 revenue. For comparison, the four largest hyperscalers have been spending in the region of 30% to 40% of cloud revenue on capital expenditure. Oracle, on guidance, will spend more than it earns.

Management frames fiscal 2027 and fiscal 2028 as peak capital expenditure years, after which the programme normalises as the capacity is delivered and the revenue catches up. That is a coherent story and it is the story the entire investment case rests on. Two things are worth testing against it.

The first is the difference between gross and net. Oracle guides gross capital expenditure of $90 to $95 billion and net cash capital expenditure of no more than $70 billion. The gap is customer prepayments and customer-supplied hardware. This is a genuine reduction in Oracle's funding requirement and it is why the company can guide to $40 billion of financing in fiscal 2027 rather than $100 billion. But it also means that a growing share of the capacity Oracle operates is not on Oracle's balance sheet, and therefore that a growing share of the revenue it records from that capacity carries a thinner margin. The company does not disclose the split between owned and customer-supplied capacity.

The second is the definition of "peak". Peak capital expenditure in absolute dollars is one thing. Capital expenditure as a share of revenue is another. If the programme plateaus at $90 billion a year while revenue grows from $90 billion to $140 billion, the intensity ratio falls from 100% to 64% and the story holds. If revenue growth stalls at $110 billion while the programme stays at $90 billion, the ratio stays above 80% and the funding requirement does not improve. Management has declined to rule out a further increase in fiscal 2028. In our base case we assume the programme plateaus in fiscal 2028 and declines modestly thereafter as a share of revenue, not in absolute terms.

There is a further consideration that receives too little attention: the capital programme is being executed at the same time as every other hyperscaler is executing an identical programme. Power, land, transformers, turbines, construction labour, high-bandwidth memory and advanced packaging are all in contemporaneous shortage. Oracle added roughly $20 billion to its 2026 capital estimate for memory costs alone. A build-out of this scale, executed in a competitive market for inputs, has a structural tendency to overrun. Oracle's disclosure of $6.247 billion of unpaid capital expenditure at quarter end, against $4.010 billion a year earlier, is consistent with a programme that is running slightly ahead of its payment cycle.

12 — Free cash flowFrom $11.8 billion of surplus to a $28.7 billion deficit in two years

Oracle's free cash flow record over the last five periods is the cleanest available summary of the change in the company's financial character. In fiscal 2024 Oracle generated $11.807 billion of free cash flow on $18.673 billion of operating cash flow and $6.866 billion of capital expenditure. In fiscal 2026 it consumed $23.686 billion. On a trailing twelve-month basis the deficit is $28.720 billion, and in the September quarter alone it was $5.396 billion.

S&P Global Ratings projects a fiscal 2027 free cash flow deficit of approximately $42 billion. Our own arithmetic is consistent with that: operating cash flow in the region of $50 billion against gross capital expenditure of $92.5 billion. The difference between the two figures and the reported quarterly numbers is the customer prepayment line, which S&P's definition treats as operating cash flow, as Oracle reports it.

A deficit of this size has exactly two funding sources: debt and equity. Oracle has used both, and section 13 sets out the detail. The relevant question for the equity holder is not whether Oracle can fund the deficit — it plainly can, having raised $43 billion of debt and $5 billion of equity in fiscal 2026 and completed a $20 billion equity issuance in the September quarter. The question is at what price, and to whose benefit.

There is a version of the Oracle story in which the deficit is exactly what a rational management team should run. A company with $664 billion of contracted demand, a cost of debt materially below the expected return on the assets, and a cost of equity that it can access should borrow and issue to build as fast as it can. The mathematics of that argument are sound. The condition on which it depends is that the expected return on the assets is in fact above the cost of capital, and Oracle does not disclose enough for an outsider to verify it.

Return on invested capital, as computed by third parties, is approximately 12.7%. Oracle's weighted average cost of capital is commonly estimated at roughly 9% to 11%, depending on the assumed cost of equity. On those numbers the spread is positive but thin — somewhere between one and four percentage points. Return on equity is much higher at 58.6%, but that is an artefact of leverage on a compressed equity base, not evidence of superior returns on the assets. The honest statement is that Oracle is currently earning a modest spread over its cost of capital while deploying capital at the fastest rate in its history, and that the spread is not large enough to absorb a material deterioration in either utilisation or pricing.

13 — Balance sheet & financing$125 billion of debt, $30 billion of leases, and a capital structure that was rebuilt in eighteen months

Oracle's balance sheet at 31 August 2026 reported total assets of $303.259 billion against total stockholders' equity of $67.196 billion. The asset base has grown by 80% in five quarters. The equity base has grown by 174% over the same period, almost entirely because Oracle sold stock rather than because it retained earnings.

Financial debt — notes payable and other borrowings, current and non-current — stood at $125.337 billion. That is composed of $7.625 billion due within twelve months and $117.712 billion beyond. The peak was $134.605 billion at 28 February 2026; two quarters of repayment have reduced it by $9.3 billion, and the company has stated that it does not expect to issue additional debt in calendar year 2026. Cash and marketable securities were $37.077 billion, so net financial debt was $88.260 billion.

The balance sheet also carries $30.594 billion of operating lease liabilities and $33.967 billion of corresponding right-of-use assets. These are data centre leases, and they are the mechanism by which Oracle has financed a large part of the physical estate without adding to reported debt. The rating agencies treat them as debt-like, which is why the leverage figures quoted by S&P and Fitch are higher than the $125.3 billion face figure. Adding the two gives approximately $155.9 billion of debt and lease obligations against $67.2 billion of equity.

Goodwill is $62.267 billion and has been essentially unchanged for four years, reflecting the Cerner and NetSuite acquisitions. It represents 92.6% of stockholders' equity and 20.5% of total assets. That is a high ratio for a company now investing in tangible assets, and it means that a large share of the reported equity base is an accounting residual rather than a recoverable pool of capital. If the Oracle Health business were to be written down — and section 18 explains why that is a live question — the equity base would fall sharply and every leverage ratio on the balance sheet would deteriorate at once.

The composition of the asset side is the clearest evidence of the transformation. Property, plant and equipment, net, was $43.522 billion at the end of fiscal 2025 and $127.845 billion at the end of the September quarter, a tripling in five quarters. Operating lease right-of-use assets rose from $29.690 billion to $33.967 billion. Total assets rose from $168.361 billion to $303.259 billion. Oracle has added $134.9 billion of assets in fifteen months, funded by $32.8 billion of debt, $24.9 billion of equity, and $31.0 billion of operating cash flow of which a substantial portion was customer prepayments.

The 31 May 2024 debt figure is total debt as reported including the current portion. Operating lease liabilities for 2024 and 2025 are the non-current portion as disclosed on the face of the balance sheet. Source: Oracle Forms 10-K for FY2024 to FY2026 and Form 10-Q for the quarter ended 31 August 2026.

Financing in fiscal 2027 is planned at approximately $40 billion, split between debt and equity, of which the $20 billion at-the-market equity programme has already been completed. Oracle also issued $5 billion of 6.50% Series D mandatory convertible preferred stock in February 2026 as part of a $30 billion combined debt and preferred raise. The preferred pays $1,625 per share annually and converts into common stock on a fixed future date, which means it is deferred equity dilution rather than a permanent claim.

The maturity profile is a genuine strength and it deserves to be stated. Fitch noted that Oracle's maturities are staggered out to 2064 and that only a small amount — between $2.8 billion and $11 billion depending on the measurement — comes due through fiscal 2026. A company with $125 billion of debt and $37 billion of cash, facing a free cash flow deficit, does not have a refinancing cliff. It has a funding programme. Those are very different risks.

14 — The credit market's verdictA BBB− rating and a CDS spread above the 2008 crisis peak

The equity market's view of Oracle has been volatile. The credit market's view has been consistent and much more negative, and because credit investors are paid to think about downside rather than upside, their view deserves weight.

Oracle's five-year credit default swap spread reached approximately 200 basis points in late July 2026. That is $2 per year to insure $100 of Oracle debt for five years, against roughly $0.53 for the average investment-grade borrower. It briefly exceeded the 197.8 basis points Oracle reached at the peak of the 2008 financial crisis. S&P Global Ratings downgraded the company from BBB to BBB− on 9 July 2026, one notch above speculative grade, explicitly citing the scale of borrowing, negative free cash flow, and OpenAI concentration.

Two interpretations of that spread are possible and both are partly right. The first is that it is a company-specific credit signal: Oracle has borrowed aggressively against contracts whose counterparties are not investment grade themselves, and the market is pricing that. The second is that Oracle has become the liquid proxy for the entire AI infrastructure complex. Oracle accounts for roughly $6.5 billion of the $12.5 billion of net credit default swap notional across major technology single names. When an investor wants to express a view on whether AI capital expenditure gets repaid, Oracle is the instrument they trade, because it is the name where the answer is most concentrated and most uncertain.

The trend since July has been more constructive. The spread has narrowed from its peak, the company has repaid $9.3 billion of debt in two quarters, and management has stated that it does not expect to issue additional debt in calendar 2026. The completion of the $20 billion equity issuance in the September quarter, while dilutive, removed a known overhang and replaced debt capacity with equity. Both moves are consistent with a management team that understands the credit signal and is responding to it. That is worth something.

15 — Equity issuance & dilution144 million new shares in one quarter, and a preferred that converts later

Oracle's share count was 2,807 million at 31 May 2025. It was 3,024 million at 31 August 2026. That is an increase of 7.7% in five quarters, and 144 million of the 217 million new shares arrived in the most recent quarter when the company completed its $20 billion at-the-market equity programme.

The at-the-market programme raised $19.909 billion of net proceeds. The average issue price was therefore approximately $138 per share, which is notable: Oracle issued equity at a price close to where the stock trades today and roughly 57% below its 52-week high. A company that believed its own equity was materially undervalued would have funded the programme with debt instead, particularly with the balance sheet capacity to do so. Issuing equity at these levels is evidence that management prioritises balance sheet resilience over near-term per-share economics, which is defensible, and it is also evidence that management did not consider the current price a bargain.

There is a second layer of dilution that has not yet arrived. Oracle issued $4.954 billion of 6.50% Series D mandatory convertible preferred stock in February 2026. Fifty million preferred shares are outstanding. A mandatory convertible must convert into common stock on a fixed date and at a fixed ratio, so the dilution is certain rather than contingent — only its magnitude depends on the conversion price. The preferred carries a $1,625 per share annual dividend, which at 50 million shares is roughly $0.32 billion a year, and $81 million was paid in the September quarter.

The cumulative effect is straightforward to state. Diluted weighted-average shares were 2,909 million in the September 2025 quarter and 3,000 million in the September 2026 quarter, an increase of 3.1%. Over the same period non-GAAP earnings per share grew 30.6%, from $1.47 to $1.92. Had the share count been held constant, the same operating improvement would have produced growth of approximately 33%. Dilution consumed roughly one tenth of it. In fiscal 2028, when the preferred converts and further equity may be issued, we model diluted shares at 3,100 million and, in the bear case, 3,150 million.

The counter-argument is that issuing equity at a price below intrinsic value is only dilutive if the proceeds are invested below the cost of equity. If Oracle earns a return above its cost of equity on the assets bought with the $19.9 billion, the issuance creates value despite the dilution. That is the whole question of the capital cycle restated in per-share terms, and it will be answered by the fiscal 2029 and fiscal 2030 returns on the assets being built now.

16 — Depreciation & useful lifeThe charge that has not been taken yet is larger than the profit that has been reported

This section contains the single most important analytical point in this report, and it is arithmetic rather than opinion.

Oracle depreciates its property, plant and equipment — which the fiscal 2026 Form 10-K describes as "comprised primarily of servers and networking equipment" — over an estimated useful life of six years. Quarterly depreciation expense was $1.351 billion in the September 2025 quarter and $3.156 billion in the September 2026 quarter, an increase of 134%. Annualised, that is roughly $12.6 billion.

Against that, gross capital expenditure is guided to $92.5 billion in fiscal 2027. If Oracle sustains a programme of that size — and management describes fiscal 2027 and 2028 as peak years rather than a one-off — then in steady state the annual depreciation charge on the resulting asset base approaches the annual capital expenditure. A six-year average life means that a steady-state programme of $92.5 billion a year produces a depreciation charge of approximately $92.5 billion a year once the base is fully seasoned.

Three qualifications are necessary and they are not small. First, this is a steady-state calculation; the charge arrives gradually over six years rather than at once, and revenue is growing over the same period. Second, Oracle's contracts are long-dated and indexed to capacity rather than to a fixed price, so revenue should grow with the asset base in a way that a one-off capital project would not. Third, the useful life assumption itself is a judgement: management has noted that four-year-old GPUs renewed at a 20% premium, which is evidence that the economic life of AI compute may be longer than the accounting life. If the economic life is eight years rather than six, the steady-state charge falls by a quarter.

What the arithmetic does establish is that Oracle's reported earnings per share materially overstate the economic earnings of the business during a period of rapid asset growth. The gap is quantifiable in order of magnitude. Capital expenditure of $92.5 billion against depreciation of roughly $14 billion means approximately $78.5 billion of capital deployed in a year that will eventually generate depreciation but does not yet. Spread over a six-year life, that is roughly $13 billion a year of deferred depreciation — about $4.30 per share pre-tax, or approximately $3.40 per share after tax at a 20% rate. On guided fiscal 2027 non-GAAP earnings of $8.10, that implies a cash-adjusted figure closer to $4.70 and a multiple closer to 30 times than the 17 times the headline figure implies.

This is not an accounting criticism. Deferring depreciation during a build-out is exactly what the matching principle requires, because the asset is not yet generating the revenue it was built to serve. It is a valuation observation. An investor buying Oracle at 17 times guided fiscal 2027 non-GAAP earnings is not buying a 17-times earnings stream. They are buying a business whose depreciation is going to rise faster than its revenue for the next four years.

17 — Quality of earningsFour adjustments that reconcile the reported profit to the cash

Reported profit and cash generation have diverged sharply, and the reconciliation between them is the most useful tool an investor has for understanding what is happening.

The first adjustment is customer prepayments, covered in section 10. Of $23.103 billion of operating cash flow in the September quarter, $11.363 billion was customer money received in advance. Core operating cash flow was $11.7 billion. This is the largest single adjustment and the one that most changes the interpretation of the reported record.

The second is the composition of non-GAAP earnings. In fiscal 2026 Oracle reported non-GAAP earnings per share of $7.63 and $6.83 excluding one-time investment gains — a $0.80 difference attributable to the sale of the Ampere chip business and warrants in Bloom Energy. The company's own fiscal 2027 guidance of $8.10 is stated as 18% growth after adjusting for those one-time events. Guidance that requires a footnote to explain why the growth rate is not what the arithmetic says is guidance that should be read with the footnote attached.

The third is capitalised versus expensed cost. The cloud and software cost of revenue line rose 77% to $6.400 billion in the September quarter against cloud revenue growth of 62%. That cost line contains the operating expense of running the data centres — power, bandwidth, labour, maintenance — but not the depreciation on the buildings and equipment, which sits below the operating expense line in a separate caption. Growth in that line faster than cloud revenue means the direct cost of serving the revenue is rising faster than the revenue. Combined with a depreciation charge that is about to accelerate, the direction of travel on infrastructure profitability is negative.

The fourth is stock-based compensation, approximately $4.5 billion in fiscal 2026 and $1.127 billion in the September quarter. Oracle excludes it from non-GAAP earnings. It is a real cost — it transfers value from shareholders to employees — and at 5.8% of revenue it is larger than the company's entire research and development budget for its applications business. We do not add it back when assessing the economics of the enterprise, though we use non-GAAP figures where they are the basis of guidance and consensus.

Read down that table and the quarter makes sense. Oracle earned $4.8 billion on a GAAP basis, $5.8 billion on its own non-GAAP basis, and generated $23.1 billion of operating cash flow of which $11.4 billion was customer prepayment and $3.4 billion was depreciation added back. It spent $28.5 billion on capital expenditure. On the cash the business actually generated from selling things, it consumed $16.8 billion.

None of those numbers is wrong and none of them is a restatement. They are four different questions asked of the same quarter, and a reader who looks at only one of them will reach a conclusion that the others contradict.

18 — Oracle HealthA $28 billion acquisition that has never earned its cost of capital

Oracle acquired Cerner for approximately $28.3 billion in June 2022. The strategic rationale was that healthcare records were a vast, under-digitised data set, that Oracle could move the estate onto its own cloud, and that a modern clinical platform would give Oracle a defensible vertical franchise for a generation. Three years later the financial record does not support the price paid.

Oracle does not break out Oracle Health as a reportable segment. It is reported inside cloud applications and services, and the company has disclosed only that the business grew at a low single-digit rate through fiscal 2026 and that its margin has been below the corporate average since acquisition. The $62.3 billion of goodwill on the balance sheet, of which Cerner is the largest component, has not been impaired. Given the trajectory of the underlying business, that is a judgement call rather than an obvious conclusion.

Management's plan is a full AI rewrite of the Cerner hospital and clinic patient care management system, described in the September release as "an all-new 100% Agentic system made up of a collection of AI agents for every medical specialty from General Medicine to Oncology to Radiology." The company expects this to push the growth rate of the overall Oracle Health business into double digits in fiscal 2027. It also announced a new AI clinical trial system designed to let regulators review trial results faster.

The competitive position is contested. Epic Systems holds the majority of the large US health system market and has been gaining share, not losing it. Oracle's customer base skews toward smaller hospitals and government providers, which are precisely the customers most exposed to reimbursement pressure and most constrained in their capital budgets. An AI rewrite is a credible way to differentiate on capability; it is not obviously a way to differentiate on price in a market where the incumbent has a decade of deployment experience and a customer base that does not want to migrate.

Our base case assumes Oracle Health grows at low double digits from fiscal 2027 and that the goodwill is not impaired. Our bear case assumes single-digit growth and a partial write-down. The difference between the two is worth roughly $6 per share on the equity, which is material but not decisive against the $231 range of the overall scenario spread.

19 — Competition I: the hyperscalersGrowing fastest, from the smallest base, in a market where scale is the moat

Oracle competes in cloud infrastructure against Amazon Web Services, Microsoft Azure and Google Cloud. All three are larger, all three have broader product catalogues, all three have larger developer communities, and all three are investing at comparable or greater absolute scale. Oracle's structural position is nonetheless better than its size suggests, for reasons that are specific to how AI workloads differ from general-purpose cloud workloads.

Oracle is growing infrastructure revenue faster than any of the three. That is partly a function of base: annualised infrastructure revenue of approximately $29.6 billion against an AWS base several times larger. Percentage growth from a small base is easier. But the composition of the growth is more interesting than the rate. Oracle's AI customers are buying very large, homogeneous blocks of compute — multi-hundred-megawatt campuses filled with a single generation of accelerators — and that workload does not reward the breadth of services that AWS and Azure have built. It rewards three things: the ability to deliver power and cooling at scale, the ability to procure GPUs, and the ability to price the whole thing below the alternative.

On all three, Oracle is credible. It has moved faster than the incumbents on gigawatt-scale campuses, in part because it has less legacy data centre estate to retrofit. It has secured GPU allocations from NVIDIA, including early access to the Vera Rubin generation, which management said is performing better than expected on quality, yield and performance. And on price, Oracle has a structural cost advantage that is not widely appreciated: it owns or controls some of the cheapest power and land in the United States, including a substantial position in Texas, and it has been willing to sign fifteen-to-nineteen-year leases to lock in the physical layer.

Against that, the moat is thin. General-purpose cloud has high switching costs because applications, data and identity are all entangled with the provider's managed services. An AI training cluster has almost none: the workload is a job that runs on GPUs, and if a competitor offers the same accelerators at a lower price, the job can move. Oracle's AI infrastructure revenue is therefore closer to a commodity rental business than to a platform business, and commodity rentals are priced to cost plus a modest spread. That is consistent with the low thirties gross margin management has disclosed.

20 — Competition II: neocloudsThe new entrants are the price-setters, and they are all losing money

The most important competitive development in AI infrastructure is not the hyperscalers. It is the emergence of a set of specialist GPU cloud providers — CoreWeave, Nebius, Lambda and others — whose entire business is renting accelerated compute, frequently financed with debt secured against the GPUs themselves and against contracted revenue.

These companies compete directly with Oracle's AI infrastructure business and they do so on price, because they have no other lever. They have no database franchise, no applications suite, no installed enterprise base. They have GPUs, power contracts and a sales force. Their pricing therefore tends toward the marginal cost of capital plus a thin return, and Oracle has to compete against that pricing for every incremental contract.

Oracle's advantage in that competition is that it does not need the AI infrastructure business to stand alone. It can price compute at a margin that would bankrupt a specialist, because the customer also buys database, applications and support. That cross-subsidy is real and it is Oracle's most durable competitive weapon in this market. It is also the reason the infrastructure segment's true profitability is impossible to assess from outside: the margin reported in the segment is not the margin the company earns from the relationship.

The specialists are also Oracle's customers in an odd way, and its competitors' customers. The same NVIDIA allocation, the same construction contractors and frequently the same financial sponsors appear across all of them. In a market where the marginal buyer is financed by the same capital that finances the seller, price signals are unreliable. That is a general observation about the AI infrastructure market in 2026, not an accusation against any particular company, and it is a reason to be sceptical of any forecast of stable pricing.

There is one further competitor that deserves mention: the customer's own data centre. A material share of Oracle's largest contracts are structured so that the customer supplies the hardware. In those arrangements Oracle is providing power, land, networking, operations and financing structure, not silicon. The economics of that business are closer to a real estate investment trust than to a cloud provider: lower capital intensity, lower margin, long-duration contracted revenue. Oracle does not disclose how much of its AI infrastructure revenue is of this type, and the answer matters more to the valuation than any of the competitive factors discussed above.

21 — Regulation & sovereigntySovereign cloud is a growth driver and a governance risk at the same time

Oracle's regulatory exposure runs in two directions. The first is defensive: as a vendor to governments, defence agencies and regulated industries, Oracle is subject to procurement rules, security requirements, export controls on advanced computing, and an evolving set of obligations around data residency. The second is offensive: those same requirements are a commercial opportunity, because a sovereign government that will not put its citizen data on a foreign hyperscaler's shared infrastructure will pay a premium for a dedicated one.

Oracle has leaned into the second. It operates sovereign cloud regions in the European Union and the United Kingdom, and has announced dedicated deployments for national governments. The strategic logic is sound: sovereign workloads are contractually long, politically protected from price competition, and require the kind of bespoke, isolated deployment that Oracle's architecture handles well. The margin on sovereign infrastructure should be materially above the margin on commercial AI rentals, for the same reason that a dedicated facility commands a premium over a shared one.

The defensive exposure is less quantifiable but not trivial. Export controls on advanced accelerators shape which customers Oracle can serve and where. The large AI contracts that transformed the backlog are concentrated in a small number of US-domiciled counterparties, and the structure of those contracts — prepayment for hardware, customer-supplied equipment, cross-border leasing — creates tax and transfer-pricing questions that will attract scrutiny. Oracle's income taxes payable rose from $11.771 billion to $12.060 billion in the quarter, and the provision for income taxes was $847 million on $5.607 billion of pre-tax income, a 15.1% effective rate. That is a low rate for a company with this much US-domiciled infrastructure, and it depends on the geographic allocation of profit.

There is also a political dimension that is genuinely difficult to assess. Data centre development has become a contested local issue in the United States. Opposition to large campuses on grounds of power consumption, water use, land use and local tax treatment has become a live electoral issue ahead of the midterm elections. Project Jupiter in New Mexico — discussed in section 22 — has faced precisely this combination of local opposition and environmental challenge. A company whose capital plan depends on siting more than four gigawatts of new capacity in the next three years is exposed to a permitting environment that has become less predictable than it was.

22 — Power, land & physicsThe constraint is no longer silicon, and the Project Jupiter notice shows why

The binding constraint on AI infrastructure in 2026 is not GPUs. It is electricity, and behind electricity, the permits, pipelines, transformers, turbines and transmission capacity needed to deliver it. Oracle's capital programme is therefore as much an energy and construction programme as a technology programme, and it should be assessed as one.

The company's disclosed positions are large. Project Jupiter in New Mexico is planned at more than 2.4 gigawatts and is to be powered by Bloom Energy fuel cells. Abilene, in Texas, has delivered 618 megawatts of a campus that totals roughly 820 megawatts. Shackelford is described as the next gigawatt-scale campus. Oracle's Stargate commitments contemplate 4.5 gigawatts. For scale, 4.5 gigawatts is approximately the output of four large nuclear reactors, and it must be sited, permitted, fuelled and connected within a few years.

On 24 September 2026, Bloomberg reported and CNBC confirmed that Oracle had sent a force majeure notice to the developer of Project Jupiter, a unit of Blue Owl Capital. The notice is designed to protect Oracle from higher expenses by allowing it to delay payment if the campus does not come online as expected in 2028. Oracle said in a statement that "Project Jupiter remains on our planned schedule" and that it is "fully committed to New Mexico." Blue Owl said the notice "does not change the financial commitments to this multi-year project." The shares fell more than 3% on the news, and Oracle's $18 billion of debt tied to the project is reported to be trading at stressed levels.

Co-CEO Clay Magouyrk told analysts on the September earnings call that Project Jupiter would not affect fiscal 2027 revenue or earnings guidance. That is almost certainly true, because the campus was not scheduled to contribute in fiscal 2027. The relevant question is fiscal 2028 and beyond, and it is the question that management declined to address.

There is a broader point about the physical layer that favours Oracle. Power and land are long-duration assets with high barriers to entry, and Oracle has been willing to commit to fifteen-to-nineteen-year leases to secure them. If AI compute demand persists, those positions appreciate. If it does not, they become fixed obligations against a stranded asset. The fifteen-to-nineteen-year lease against a five-year customer contract, noted by S&P in its downgrade, is the clearest statement of the maturity mismatch at the heart of this business.

23 — ValuationA 17-times multiple on earnings that are flattered by depreciation, and a 2.8-times spread between the low and the high

Oracle trades at $139.54, a market capitalisation of $421.9 billion, and an enterprise value of approximately $515 billion after adding $88.3 billion of net financial debt and $5.0 billion of preferred stock. On trailing twelve-month GAAP earnings of $6.38 per share the multiple is 21.9 times. On company guidance of $8.10 of non-GAAP earnings per share for fiscal 2027 the forward multiple is 17.2 times. On trailing revenue the price-to-sales ratio is 5.9 times, and on guided fiscal 2027 revenue of $90 billion it is 4.7 times.

The chart above is a statement about the market's uncertainty rather than about the company's value. Within a single year the shares have traded between $114.50 and $322.54 — a spread of 2.8 times between the low and the high. The consensus target of $236.52 across 47 covering analysts sits 69.5% above the current price, which is an unusually large gap and itself a signal that the analyst community has not reconciled the operating results with the cash flow statement. We are comfortable being materially below consensus. When every covering analyst has a target 40% or more above the market price on a large-cap company with a contested capital cycle, the disagreement is about the model, not about the facts.

Our approach is deliberately conservative on the one variable that matters most. Rather than valuing Oracle on a free cash flow that will not exist for several years, we value it on estimated fiscal 2028 non-GAAP earnings per share applied to an exit multiple, and we stress the two inputs that determine the answer: the margin at which the infrastructure revenue converts, and the multiple the market will pay for a levered infrastructure operator.

Farstar estimates. Fiscal 2028 is the year ending 31 May 2028. Non-GAAP earnings per share assumes 3,050 million diluted shares in the bear case, 3,100 million in the base case and 3,150 million in the bull case, reflecting the $4.954 billion mandatory convertible preferred and assumed further equity issuance. Interest expense is modelled at $9.5 billion in the bear case, $8.5 billion in the base case and $8.0 billion in the bull case. A 20% tax rate is applied throughout. No terminal value is used; the multiple is the exit multiple.

The bear case is not a disaster scenario. It assumes revenue grows from a guided $90 billion in fiscal 2027 to $98 billion in fiscal 2028, which is 9% growth — slower than any year since fiscal 2020, but not a contraction. It assumes non-GAAP operating margin compresses from 41.6% to 29% as depreciation on the September 2025 to fiscal 2027 capital programme arrives. And it assumes the market applies a 12-times multiple, which is roughly where levered infrastructure and utility businesses trade. The result is $59, or 58% below the current price.

The base case assumes revenue of $117 billion in fiscal 2028, 30% growth, and non-GAAP operating margin of 38% — a compression of 3.6 percentage points from the fiscal 2027 level, which is what we believe depreciation and mix will cost. Earnings per share of $9.20 on 3,100 million shares implies net income of $28.5 billion, which is consistent with $44.5 billion of non-GAAP operating income, $8.5 billion of interest and a 20% tax rate. An 18-times multiple gives $166.

The bull case assumes revenue of $134 billion in fiscal 2028 — 49% growth — and margin of 43%, which requires the multicloud database business to scale faster than we assume and the AI infrastructure margin to hold above the low thirties. It assumes a 23-times multiple, which is a growth-company multiple. The result is $290.

Weighted at 30/50/20 the value is $159. We set a 12-month target of $160, approximately 15% above the price, and we rate the shares Neutral.

24 — The bull caseSix arguments we take seriously

We present the bull case at full strength. If we cannot state the strongest version of the opposing argument, our own position is not worth publishing.

• The backlog is contracted, not forecast, and it is diversifying. $664 billion of signed commitments against $71.8 billion of trailing revenue is 9.2 times coverage, and the non-OpenAI book more than doubled over the past year. Oracle booked more than $30 billion of new AI contracts in a single quarter while recognising only $7.4 billion of infrastructure revenue. Even on a 12%-in-twelve-months conversion assumption, the near-term revenue is contractually protected in a way that almost no other large-cap company's is.
• The customers are funding the build. Prepaid and customer-supplied hardware across the large AI contracts totals $75 billion. Net cash capital expenditure is guided to no more than $70 billion against gross capital expenditure of $90 to $95 billion, and the $20 billion at-the-market equity programme was completed without difficulty. Oracle is acting as a project finance vehicle for its customers' AI ambitions rather than betting its own balance sheet on speculative workloads, and the debt level reflects that.
• Utilisation is 97.9% and four-year-old GPUs are renewing at a premium. This is the strongest available evidence that the assets being built are not speculative and that the economic life of AI compute is longer than the accounting life. If the useful life is eight years rather than six, the steady-state depreciation burden falls by a quarter and the entire bear case weakens materially.
• Operating leverage is real and visible. Non-GAAP operating margin held at 42.1% while infrastructure revenue grew 121%. GAAP operating income grew 57% on revenue growth of 30%. Sales and marketing fell 12%, research and development fell 4%, and amortisation fell 52%. A company buying growth cannot produce that combination. Whatever the mix does to gross margin, the operating line is being managed with discipline.
• The multicloud database is the most valuable undisclosed asset in the company. Growing 353% with 60% to 80% gross margins and essentially no incremental capital, it converts Oracle's biggest competitive liability — that running Oracle Database meant committing to Oracle Cloud — into a distribution advantage inside three competitors' data centres. Every dollar of it offsets the mix drag from infrastructure.
• The price now embeds a great deal of bad news. At 17.2 times guided fiscal 2027 non-GAAP earnings, 4.7 times guided revenue, and 56.7% below the 52-week high, the shares no longer discount a smooth execution. The company has repaid $9.3 billion of debt in two quarters, stated that it will not issue additional debt in calendar 2026, and holds $37.1 billion of cash. The asymmetry has improved materially from where it was twelve months ago.

25 — The bear caseSix arguments that keep us at Neutral rather than at Buy

We have set out the bull case at length because it is a real one, and because the market has spent twelve months pricing it. What follows is the other side, and it is the side we find more persuasive at $139.54. None of these six points is a prediction that the company fails. Each is a reason to believe that the earnings the shares are currently capitalised against are of lower quality, and arrive later, than the income statement implies.

• The backlog is a claim on somebody else's balance sheet, not on Oracle's. The $664 billion of remaining performance obligations is a genuine contractual commitment and we do not treat it as fiction. But roughly half of it is attributed in press reporting to a single counterparty, OpenAI, which lost $20.9 billion in 2025, raised $110 billion in February 2026 from Amazon, NVIDIA and SoftBank, and whose ability to fund a five-year, roughly $300 billion compute commitment depends on raising several times more capital than it has raised to date. Oracle has contracted to build 4.5 gigawatts for that counterparty. The backlog converts into revenue only if the counterparty keeps paying, and the revenue is recognised on delivery of capacity that Oracle must fund in advance with its own and its lenders' money. We are not calling the counterparty a credit event. We are observing that Oracle has taken a supply obligation that is largely uncollateralised and concentrated, and has booked it as an asset.
• The record operating cash flow is one third financing. The headline number for the August quarter was $23.103 billion of net cash from operating activities, up 184% year over year. That figure includes an $11.363 billion increase in deferred revenues arising from "customer prepayments with significant financing component" — a line that contributed nothing in the comparable quarter a year earlier. Strip it out and core operating cash flow was $11.740 billion against $28.499 billion of capital expenditure: a real deficit of $16.759 billion in a single quarter. The company disclosed the financing nature of the prepayments, which is to its credit, but the cash flow statement does not separate them, and the headline is the number that gets quoted. A reader who takes operating cash flow at face value will conclude the business is funding its own build. It is not.
• Depreciation has been deferred, not avoided. Oracle capitalised $55.663 billion in fiscal 2026 and is guiding to $90–95 billion in fiscal 2027, against $3.156 billion of quarterly depreciation in the August quarter — a run rate of roughly $14 billion a year. Servers are depreciated over a six-year useful life, and the FY2026 Form 10-K states that assumption explicitly. A $92.5 billion annual capital programme, held for six years, implies a steady-state depreciation charge in the region of $15 billion a year arriving on a lag of two to three years. After tax at the company's 15% effective rate, the incremental charge is worth roughly $3.40 per share against a guided fiscal 2027 non-GAAP EPS of $8.10. That charge is not in current earnings. The multiple that looks like 17 times is closer to 30 times on a cash-adjusted basis. If the company were to shorten the assumed useful life — the honest response to a generation of accelerators that are physically and economically obsolete well before six years — the charge arrives sooner.
• The credit market has already repriced the story. S&P downgraded Oracle to BBB− on 9 July 2026, citing the scale of borrowing, negative free cash flow and concentration with OpenAI. Five-year credit default swap spreads on Oracle reached approximately 200 basis points in late July — above the 197.8 basis points touched at the peak of the 2008 financial crisis, against 53 basis points for the investment-grade index, 93 for Meta and 78 for NVIDIA. Oracle accounts for roughly $6.5 billion of the $12.5 billion of net CDS notional written on technology companies. Equity investors are being asked to buy a 17-times forward multiple while the debt market prices the same cash flows at a spread that implies a materially higher probability of impairment. One of the two is wrong.
• The assets are longer than the contracts that pay for them. Oracle's data-centre leases run fifteen to nineteen years. Its customer commitments with the largest AI counterparties run approximately five years, and the company's own disclosure of the conversion schedule shows only about 12% of the backlog converting inside twelve months and a further 34% in the following twenty-four. The company has mitigated this in places by requiring customer-supplied hardware and prepayments, and the $11.4 billion of prepayments in the August quarter is evidence that it is negotiating hard. But a mismatch of that duration is a structural exposure, not a presentational one: it means Oracle carries residual-value risk on assets whose economic life is set by a technology cycle, against counterparties whose contractual obligation ends at five years. Project Jupiter is the illustration. On 24 September 2026 Oracle served a force majeure notice on the developer of the New Mexico campus, seeking to defer payment obligations if the site misses its 2028 online date. Approximately $18 billion of project debt was reported trading at stressed levels. The co-chief executive said the notice would not affect fiscal 2027 guidance, which is plausible for a single quarter and less so as a general principle.
• Equity has become the funding source of last resort. Oracle raised $43 billion of debt and $5 billion of equity in fiscal 2026, and completed a $20 billion at-the-market equity programme in the August quarter for $19.909 billion of net proceeds — an average price that a 52-week range of $114.50 to $322.54 suggests was well below where the shares traded when the programme was authorised. Share count went from 2,880 million at 12 June 2026 to 3,024 million at 31 August 2026, a 5.0% increase in one quarter. S&P projects a fiscal 2027 free cash flow deficit of approximately $42 billion and dilution of roughly 4.8% from a further equity raise. A company that must sell equity to fund operations has its cost of capital set by its share price, and its share price is set by the confidence that it will not have to sell equity. That is a reflexive loop, and it is the same loop that closed on the digital-asset treasuries earlier in this cycle.

26 — Risk matrixTwelve risks, scored by probability and impact

We score each risk on probability (1 = remote, 5 = already happening) and impact on intrinsic value per share (1 = immaterial, 5 = changes the investment case). The product is the score. Five risks score at or above 15 and are marked critical; in our judgement they are the five that decide whether the equity is worth $59 or $290.

Cells read left to right as impact rises from 1 to 5, top to bottom as probability rises from 1 to 5. Shading is the probability–impact product: green below 8, amber 8 to 11, red at 12 and above. Superscripts identify risks from the register above.

Two observations about the shape of this matrix. The first is that the register is bottom-heavy: five of twelve risks score critical, and they are not independent of one another. R2, R3 and R6 are all manifestations of the same underlying question — whether the capital markets will keep funding a build whose returns arrive in the 2030s. R4 and R5 are both consequences of the same physical fact, which is that the assets Oracle is buying take years to arrive and years more to earn out. A portfolio of risks that correlates this heavily should be underwritten as one risk with a wider distribution, not as five separate line items with individual probabilities.

The second is that the two risks we score lowest are the two the market discusses most. Regulatory intervention (R9) and dilution (R12) both score 6. Dilution is real and we model it — S&P's 4.8% estimate is reasonable — but it is a second-order effect against a share price that has already halved, and it is disclosed and managed. Regulation is a genuine consideration in Europe and a genuine driver in sovereign markets; it has not yet been the binding constraint, and the binding constraint is electricity and capital.

27 — CatalystsWhat we are watching, and when

The next four months contain more information than the last four quarters did. Oracle reports its December quarter on 10 December 2026, the company holds its annual customer event in Las Vegas from 25 to 28 October, and at least one credit rating committee will have to decide whether BBB− with a negative outlook is a stable resting place. Each of the following is capable of moving the shares by more than 10% on its own.

28 — ConclusionNeutral at $160: an excellent business attached to a financing structure that has not been tested

Oracle in 2026 is two companies wearing one ticker. The first is the one that has existed for four decades: a database franchise with extraordinary switching costs, an applications portfolio that generates high-margin recurring revenue at a scale few competitors match, and a customer base that has renewed with Oracle through every technology transition since the mainframe. That company earned $28.9 billion of non-GAAP operating income in fiscal 2026 on $67.4 billion of revenue, and it would be worth a great deal on its own. The second is a company that did not exist three years ago: a builder and lessor of gigawatt-scale AI infrastructure, funded with debt and equity raised against a backlog that is concentrated in one counterparty and denominated in megawatts rather than cash.

The bull case is that the second company is the first company's salvation. Oracle had lost the cloud transition; the arrival of generative AI created a demand shock so large that a late entrant with capital and power contracts could win share anyway, and the $664 billion backlog is the evidence that it did. Cloud revenue grew 62% in the August quarter, infrastructure revenue grew 121%, the company delivered more than 300,000 GPUs and 850 megawatts of new capacity, and the multicloud database has turned the oldest liability in the business into distribution inside the three clouds that were supposed to kill it. Those are not promotional claims. They are reported results, and they are impressive.

The bear case is that the second company is consuming the first. Free cash flow was negative $23.7 billion in fiscal 2026 and negative $5.4 billion in the August quarter. Notes payable rose from $92.6 billion to $125.3 billion in twelve months and peaked at $134.6 billion in February 2026. The equity was sold into the market at prices far below the 52-week high, and the credit market repriced the name above its 2008 crisis spread. The depreciation that would make the infrastructure business look like the capital business it is has not yet been recognised. And the counterparty that accounts for half the backlog is a company that loses $20.9 billion a year and depends on the same capital markets that Oracle depends on.

Our judgement is that both are true, and that the price is somewhere between them. At $139.54 the shares trade at 17.2 times guided fiscal 2027 non-GAAP earnings and 4.7 times guided revenue, 56.7% below the 52-week high of $322.54. That is not a valuation that assumes perfection; it is a valuation that assumes the plan roughly works. Against a consensus target of $236.52 from 47 analysts, and a base case of our own at $166, the arithmetic offers some upside. Against a bear case at $59 — a 12-times multiple on $4.95 of earnings in a world where the backlog converts slowly and the depreciation arrives early — the downside is severe. Weighted at 30/50/20, the expected value is $159, marginally above the current price. That is not a sufficient margin of safety for a business whose cash flows are one third financing and whose largest customer is unprofitable.

What would change our mind, in order of importance. A quarter in which operating cash flow excluding customer prepayments exceeded capital expenditure would be decisive: it would mean the infrastructure is earning its cost of capital rather than consuming it, and it would justify a materially higher multiple. A disclosure that separates the financing component of prepayments from operating cash flow in the body of the cash flow statement — rather than in the notes — would tell us management wants the quality of the number understood, which is a signal about everything else. And a depreciation policy that reflects the economic life of AI accelerators, with the earnings consequence accepted rather than deferred, would remove the largest single source of disagreement between our model and the reported accounts.

Conversely, three developments would move us to a sell rating. A downgrade to sub-investment grade, which would raise the cost of refinancing $7.6 billion of current maturities and reprice the entire stack. A material reduction in the OpenAI commitment, or evidence that the counterparty is renegotiating delivery schedules rather than volumes. And a fiscal 2027 outturn in which capital expenditure exceeds $95 billion while cloud revenue growth decelerates below 40% — the combination that would confirm the bear case's central claim, which is that Oracle is buying growth it cannot afford at a price it cannot recover.

The next two quarters will resolve most of this. We will re-cut the report after the December quarter and, if the depreciation question is addressed in the fiscal 2027 filings, will treat that as a full revision rather than an update.

29 — Appendix: financialsReported figures, and the ones we derived

The tables below consolidate the primary-source figures used throughout this report. Where a number is a Farstar derivation rather than a company disclosure, it is marked with a dagger and the method is stated in the notes. All company data is drawn from Oracle's filings with the US Securities and Exchange Commission and from its earnings releases for the periods indicated.

^† Derived. Fiscal 2025 cloud revenue and its components are derived by applying the fiscal 2026 growth rates disclosed by the company (39% total cloud, 77% infrastructure, 11% applications) to the fiscal 2026 figures; they are presented for comparability and will not reconcile exactly to the fiscal 2025 release. First-quarter applications revenue is derived by subtracting disclosed infrastructure revenue from disclosed total cloud revenue. Core operating cash flow is reported operating cash flow less the increase in deferred revenues arising from customer prepayments with a significant financing component. Free cash flow is operating cash flow less capital expenditure and is not a company-reported measure.

Notes payable at the first-quarter date comprise $7.625 billion of current and $117.712 billion of non-current obligations, a total of $125.337 billion. Operating lease liabilities of $30.594 billion are presented separately and are not included in notes payable; a reader who consolidates them arrives at a total obligation of approximately $155.9 billion, which is the basis on which several third-party data providers report Oracle's debt. Total debt peaked at $134.605 billion on 28 February 2026 and has been reduced in each subsequent quarter. Common shares outstanding at 12 June 2026 were 2,880,471,000 per the FY2026 Form 10-K; the first-quarter figure is the count used in the diluted earnings per share calculation.

^†† The fiscal 2027 first-quarter figure of $664 billion was reported as an increase of $209 billion year over year and $26 billion sequentially. The step change between the first and second quarters of fiscal 2026 reflects the signature of the large AI infrastructure contracts in that period and is the single most important discontinuity in Oracle's financial history.

30 — SourcesPrimary and secondary references

• Oracle Corporation, Form 10-Q for the quarter ended 31 August 2026, filed with the US Securities and Exchange Commission, including the consolidated balance sheet, statement of operations and statement of cash flows.
• Oracle Corporation, Form 10-K for the fiscal year ended 31 May 2026, including the property, plant and equipment note, the debt note, the leases note and the revenue-recognition disclosure of remaining performance obligations.
• Oracle Corporation, quarterly earnings releases for the fourth quarter of fiscal 2026 (10 June 2026) and the first quarter of fiscal 2027 (10 September 2026), and the accompanying non-GAAP reconciliations.
• Oracle Corporation, XBRL company facts as filed with the SEC, CIK 0001341439, used for the historical revenue, backlog, capital expenditure and debt series reproduced in the appendix.
• Oracle Corporation, fourth-quarter and first-quarter earnings call transcripts, including the prepared remarks of the co-chief executives and the chief financial officer, and the question-and-answer session.
• Oracle Corporation, Form 8-K filings relating to the $20 billion at-the-market equity offering programme and the senior notes issuances and repayments of fiscal 2026 and the first quarter of fiscal 2027.
• S&P Global Ratings, research update on Oracle Corporation dated 9 July 2026, downgrading the issuer credit rating to BBB− and discussing leverage, free cash flow and counterparty concentration.
• Credit default swap spread data for Oracle Corporation, Meta Platforms and NVIDIA, and the Markit CDX investment-grade index, as reported in market commentary in July 2026.
• Press reporting on the Project Jupiter campus in New Mexico, the force majeure notice served on its developer on 24 September 2026, and the trading levels of the associated project debt.
• Press reporting and public disclosures relating to OpenAI, including its 2025 loss of $20.9 billion, the $110 billion financing round of February 2026, and the five-year compute agreement with Oracle.
• Market data for Oracle common stock, including the closing price of $139.54 on 24 September 2026, the 52-week range of $114.50 to $322.54, the market capitalisation of $421.93 billion and the consensus twelve-month price target of $236.52 compiled from 47 contributing analysts.
• Third-party commentary on the useful life of AI accelerators, on the economics of neocloud operators including CoreWeave and Nebius, and on the structure of gigawatt-scale data-centre leases.

Where this report cites a figure that is not a company disclosure — a Farstar estimate, a derived quarter, a scenario valuation or a cash-adjusted multiple — it is labelled as such at the point of use. Market data is as of the close on 24 September 2026 unless otherwise stated.

31 — DisclosurePosition, basis of preparation and limitations